Working with AI

Write a useful request

Include the goal, scope, constraints, and acceptance criteria. For complex work, ask AI to investigate and propose a plan before editing.

Add files, images, code locations, or an Appshot to the composer when they clarify the task. Type / to select an available Skill.

After you send a local image, Wework keeps its preview in the message. The image remains available when you reopen Wework or return to the conversation after switching away. If the original local file is deleted, the preview cannot be restored.

Voice input

Select the microphone at the bottom of the task composer to start voice input. On first use, follow the prompts to prepare the speech model and grant system microphone permission. You can also prepare the model under Settings → General → Voice input.

Start with INT8 (lightweight) for everyday dictation to reduce download size and memory use. If it makes frequent recognition errors, try FP32 (full precision), which retains full 32-bit floating-point precision with a larger download and higher memory use. Both the settings page and the first-use setup dialog explain these choices.

While voice input is requesting microphone access, recording, transcribing, or showing feedback, other actions on the left of the composer are hidden and voice controls remain available. Finishing or cancelling restores those actions. Switching conversations also exits voice input.

Stop recording to transcribe. The recognized text is inserted at the position captured when recording started, ready to review, edit, and send. Cancelling leaves the draft unchanged. If you edit the draft during recording, the transcript is retained; select Insert text to insert it at the current cursor position.

Voice input reuses the official DSH plugin. Once downloaded, the default local recognition model transcribes on the device.

Edit Markdown and tables

Pasted Markdown appears as editable content in the composer. Sent messages also render tables, lists, bold text, and code. Use Shift+Enter for a new line and Tab / Shift+Tab to move between table cells.

Hover over a message table to copy its Markdown or expand it; press Esc to close the expanded view. Click an HTTP/HTTPS link in the composer to open its actions. Sent links open using your browser preference. Text appended after a pasted link stays separate and does not change its destination.

Repair a PR from the task sidebar

Click the PR/MR status beside a task and choose Continue fixing with AI when an abnormal status is shown. For a running task, the repair request guides the current response. For an idle task, it starts the next turn in the same conversation. If the current response finishes just after you click, the request is sent as the next turn. Delivery failures show an error so you can try sending again.

Add a conversation to a project-space board

Runtime task and private conversation titles are independent of the bound Issue title. Runs started from tasks, Issue details, or Collaboration never overwrite an existing Issue title when the conversation title changes. Edit the title in Issue details to rename the Issue.

After enabling Experimental features, open the composer's + menu and select Project space. Selecting a space opens the Add to "Project space name" dialog even before a new task sends its first message. Choose Create board task or Link an existing board task. Confirmation saves the choice; the first message creates the card or links the selected Issue. Cancelling keeps the previous selection. Select the current space again to change the association choice or target Issue.

The selected destination appears as Add to board · Project space name below the composer so you can confirm it before sending. A project space inherited from the local-project automatic-join setting appears through the same control and creates a new board task after the first message by default.

For an existing task, open the right-side Environment panel and select Link project space. You can link the current project or task to a local or cloud project space, or quickly create a task in that space. Local-space operations remain on the current device; cloud-space operations use shared cloud data.

When creating an issue in Collaboration, enter at least a title or content. If you enter content only, Wework preserves the complete content as the description and generates a title using the same rule as the regular task composer. A manually entered title takes precedence when both fields are present.

After selecting Let AI handle it in an issue detail, use the task composer to choose a local project, workspace or worktree, and branch before creating a local run linked to the issue. Project spaces and local code projects are separate resources. After a task is created successfully, Wework remembers the most recently used local project and workspace for that project space on the current device. The next task created from the same project space preselects that location, and a successful task created after a temporary selection change updates the mapping. The mapping does not sync to other devices.

When an issue already has linked tasks, the Linked tasks section directly shows every running task and the two most recent tasks. If more history is available, use Expand and Collapse instead of opening a separate task-list view. The action for creating another task appears beside the section heading.

After a task is linked to a board, select Change board link from the task summary above the composer, then choose the destination project space. Wework offers two choices:

• Create board task creates and links a card using the current task title and execution status.
• Link an existing board task searches editable cards in the destination board by task ID or title and links the current task to the selected card without changing that card's status.

When the destination differs from the current board, Wework asks you to confirm the move. After confirmation, the runtime task keeps only one user-selected board link. The original board card remains in place but is no longer linked to the runtime task. A newly created card enters the column that matches the current task state; for example, a settled task awaiting confirmation enters In review.

Review board-task progress

The linked-task progress area does not repeat the runtime task title. While a task is running, the first row shows the latest AI text or thinking state, and a second indented row with a short vertical guide shows the most recent tool or edit action. PR/MR status remains a trailing action. After execution stops, the card shows only the last non-empty line from the final response of the latest turn and never falls back to an older turn. Unread cards use a distinct blue-accent background and border in addition to the unread indicator, then return to the standard card style after they are opened.

Click a card to open a lightweight task workspace without leaving the board. It initially shows the latest user message and AI response, reusing the Task conversation's thinking, tool-call, and file-edit rendering. Longer conversations scroll inside the transcript area, and Load earlier history fetches older turns. The composer stays on one line until clicked, then expands with the same quick phrases and actions available in Task conversations. While the composer is active, the preview stays pinned until its top-right close button is clicked. When one board task has several tasks running, the preview initially lists a summary for each one. Selecting an individual task narrows the preview to that task's progress. To open the full Task page, use the corner action that appears at the bottom right of the currently hovered or focused card; the action does not appear on the other cards at the same time.

Create the first project and hand work to the device assistant

When you enter Collaboration with no projects yet, Wework presents one continuous getting-started path: create a project, create an issue, hand it to an assistant, then review and accept the result. Select Create project first, then choose a local or cloud collaboration space; storage location is not split into separate entry points before the project flow begins.

When an executable issue is open in Wework, its detail view shows the Device assistant. It represents the default execution environment on the current device and uses the models, plugins, Skills, local files, and desktop capabilities already available there. Selecting Hand off to Device assistant opens the task composer. The resulting run stays linked to the current issue so its result can be reviewed and accepted from the issue.

The Device assistant is not a separate cloud Agent that must be created, and it does not synchronize device accounts, files, or system permissions to other devices. To reuse or share a stable prompt, plugin, Skill, or MCP configuration, create an Agent and configure its capability source explicitly; device-specific capabilities still depend on the device that runs it.

Manage project-space automation

Project spaces provide an Automation entry that manages robot members and their execution queue in one place. A robot can be assigned as the owner of a project task: it claims the task, executes it on the local or a selected device, and writes the result or failure back into the task comment thread for human acceptance.

The execution queue shows each robot task's state (pending, queued, claimed, running, failed for retry) and supports keyword search. Completed automation runs remain as comments in the task detail, where a human can accept the result or ask the robot to continue.

Robots in a local project space can only bind to the local executor or a companion App device; cloud devices are reserved for cloud project spaces. The local App claims the queue every 3 seconds, keeps running tasks leased, and automatically requeues runs that were interrupted after their lease expired, so a run can never stay stuck in "running".

Run a cloud device on this machine (for development testing)

Robots in a cloud project space can execute on cloud devices. During local development you can register this machine's executor as a device_type=cloud device with the repo script, so the cloud-device dispatch path can be verified locally:

bash executor/scripts/dev-cloud-device.sh start    # start and keep online (idempotent)
bash executor/scripts/dev-cloud-device.sh status   # show running/online state
bash executor/scripts/dev-cloud-device.sh restart  # restart manually
bash executor/scripts/dev-cloud-device.sh stop     # stop

• The default device id is cloud-device-dev (override with DEVICE_ID); it appears under "cloud devices" in the device list after registration.
• The script builds the latest executor and watches executor/src, Cargo.toml, and Cargo.lock; source changes trigger an incremental build and dynamic executor restart. It also mints a 30-day token from the backend/.env secret and uses an isolated executor home and Codex home (it never reads personal Codex credentials).
• When creating a robot, choose the "cloud" execution environment and cloud-device-dev as the device; assigned tasks then execute locally through the cloud-device protocol and write back comments and execution records.
• Runtime data and logs live in ~/.wegent-executor-cloud-device/.

Models and devices

The model provides the AI capability; the device determines where files and commands run. Local models run on the local device. Cloud models and devices require a Wegent connection.

The coding-tool selector in the task composer chooses a local runtime such as Codex or Claude Code. The task composer currently does not provide a direct cloud Wegent Agent selector. To have a cloud Agent process project tasks automatically, configure its robot and execution environment under project-space Automation.

Open the model selector in the composer to see the account source, selection mode, current model, and reasoning effort in separate rows:

• Recommended model set: each slider position selects a model and reasoning effort combination. Moving the slider can change models. Other model sets display their own names.
• Fixed model: after you explicitly select a model, the slider changes only that model's reasoning effort. Use the reset button to restore the default recommended model set.

Click the mode or model name to open the full list, organized by local ChatGPT account, cloud models, and custom models. Long model names wrap in full, and the panel scrolls when space is limited.

Wework remembers the model, selection mode, and reasoning effort separately for new-task defaults and active tasks. Returning to the page keeps an explicitly fixed model fixed, even when it also belongs to the recommended set:

• With no task open, the composer selection belongs to the current draft and the next new task. Standalone tasks save this selection as the default for later new tasks.
• When a local project follows the global default, a project-composer selection updates the global new-task default used by later standalone tasks and other projects that follow the global default. When a project defines its own model and reasoning effort, a temporary composer change applies only to that task; starting another task restores the project default.
• After a task starts, its model and reasoning effort belong to that task. Sending a follow-up persists the latest selection, so returning to the task restores it without changing the new-task or project default.

After a conversation starts, you can still switch between official GPT/Codex models and third-party models. Wework asks for confirmation when the selected model changes, and the new model applies to the next message.

If the existing context contains encrypted reasoning or compaction state produced by the previous provider, the executor removes those non-portable fields from the switch request so the target model does not return invalid_encrypted_content. Portable messages, tool results, and other context remain available.

Control a running task

The sidebar keeps running tasks visible with their running indicators. Collapsing or expanding a project or its history list does not change execution state. When you continue a conversation, completion from an earlier turn does not end the new running turn; its indicator clears when the current turn ends.

Context compaction is shown as complete only after its turn succeeds. Plugin synchronization and idle-only process restarts wait during compaction, and a failed or interrupted compaction is not shown as successful. A response timeout does not mean compaction has stopped. A live compaction retains its activity protection until its native turn ends; use Pause to interrupt it.

While a start request remains unconfirmed, a previous completion snapshot cannot clear its running indicator. Once the executor confirms execution has started, completion of the current turn ends the running state normally.

Action	Shortcut	Use it when
Send after the response	Enter	Queue the next instruction
Steer the response	Command/Ctrl + Enter	Adjust the current direction
Interrupt and send	Command/Ctrl + Shift + Enter	Stop an incorrect direction immediately

Interrupting stops the current response but does not roll back completed file edits or commands.

Answer AI questions

Question cards show the full question above the answer field. Long questions wrap, and oversized cards scroll. Questions without options remain visible while you type. Enter your answer in the field labeled “Type your answer”, then click Submit.

Use the personal supervisor

Open the composer's + menu in a new Codex conversation or an existing Codex task and select Personal supervisor. While the main AI is working, the executor periodically reads its recent progress in the background and evaluates goal drift, missed constraints, destructive actions, and obvious blocked loops with a lightweight read-only call. Supervision is a regular feature and does not require Experimental features. It does not fork the original task, and checks continue without keeping the task view open.

• Suggest shows a correction above the composer for you to approve or dismiss.
• Auto-correct steers an active response when a clear deviation is found, or starts a normal follow-up just as if you had sent the instruction from the composer.

Supervision settings belong to the current task. Select an independent review model and a review frequency of 10 seconds, 30 seconds, 1 minute, or 5 minutes. Wework remembers the most recently saved review model and frequency on the current device and restores them when supervision is configured for a new task. Set default supervisor principles under Settings → Context; they are prefilled when supervision is first enabled and can then be customized for that task.

After supervision starts, the right-side Environment panel shows its status and the next scheduled review time. Select Review now to inspect the current progress immediately. The action is temporarily disabled while a review is already running so the task cannot start concurrent supervisor checks.

Review the processing timeline

The Processed section in an AI response displays tool calls from top to bottom by their actual creation time. Even when executor events arrive out of order, commands, file operations, and other tools created earlier remain above later activity so the timeline reflects the real execution sequence.

Completed processing activity is summarized by type. Terminal commands and Node.js REPL JavaScript calls count as commands, MCP and other capabilities count as tools, and file reads or edits use their own categories. Expand the processing area, then select a row or its chevron to inspect the details. The collapsed row uses a readable short name; the detail view shows the full tool name, invocation input, and returned output. Status-only activity without input or output does not show an empty disclosure control.

A plan generated in Plan mode remains part of the visible response instead of being placed inside the completed Processed disclosure. It stays directly visible after restoring a background task or reopening the conversation. Ordinary commands, file operations, and tool calls continue to use the summary and disclosure behavior described above.

Manage conversations

Conversations can be continued, renamed, copied to another device, or archived. Project conversations stay under their project; standalone conversations appear in the conversation list.

Right-click a started task and select Move to project to place it under another available project on the same device. The task retains its session, history, and execution directory, and remains under the destination project after restarting Wework.

Export a conversation

Select Export conversation at the top of a conversation, or open the conversation item's menu, to export the complete conversation as Markdown or HTML. Before exporting, choose whether to include the message body, tool calls and file changes, thinking, images, and other attachments. Included tool calls, file changes, and thinking are collapsed by default in HTML so the conversation remains easy to read.

• Markdown includes only the message body by default; images are not selected by default. A text-only export creates an .md file. Including images or other attachments creates a ZIP archive whose Markdown file references the packaged assets with relative paths.
• HTML includes the message body and images by default. Images are embedded so the HTML file works offline. Including other attachments that must remain separate creates a ZIP archive.

After you choose a save location, the export window continues to show packaging and writing progress. It reports completion only after the background export task finishes and the destination file is confirmed. When replacing an existing file, a failed write leaves the original file intact.

To branch from a completed AI response, move the pointer to its message actions and choose Continue in new task. The new task keeps the conversation context through that response and uses the same workspace. The source task and later messages remain unchanged. The action is hidden for streaming, cancelled, or unmappable Codex turns.

When you switch away while viewing the bottom of a conversation, Wework preserves its follow-latest state. Returning to a conversation that continued streaming in the background shows its newest content. If you scrolled upward before switching, Wework restores that reading position instead.